Automatic DFM classification method for rapid stroke diagnosis and system supporting same

Abstract

An automatic DFM classification system for stroke diagnosis according to the present invention comprises: an image data input unit for receiving the input of a DWI image and FLAIR image; a data combination unit for generating combined image data by combining the input DWI image and FLAIR image; a DFM classification modeling unit for deriving a DFM inference value from the generated combined image data by using a DFM classification model; and a DFM classification unit for performing DFM classification on the basis of the derived DFM inference value.

Description

DFM automatic classification method for rapid stroke diagnosis and supporting system

[0001] The present invention relates to a method for automatically classifying DFM, DBOM, and DAM for stroke diagnosis and a system supporting the same.

[0002] The number of stroke patients is increasing due to factors such as an aging population and Westernized dietary habits. Reperfusion treatments such as thrombolysis and mechanical endovascular thrombectomy (EVT) can significantly improve the prognosis of patients with acute stroke.

[0003] Diffusion-FLAIR (Fluid attenuated inversion recovery) mismatch, a method for distinguishing between recoverable and non-recoverable tissues in reperfusion treatment of large core stroke, reflects tissues that may still be viable among tissues that were previously assumed to be ischemic core, and thus contains the possibility of improvement due to reperfusion treatment. Therefore, determining whether there is an accurate match between diffusion-weighted imaging (DWI) and FLAIR images is used as an important indicator for determining indications for acute stroke treatment.

[0004] However, the conventional evaluation of DWI-FLAIR mismatch relies entirely on the visual judgment of medical professionals, and is highly susceptible to individual subjective judgment during the image interpretation process, which can reduce the reliability and speed of the interpretation results. In addition, there is a problem that the accuracy of the analysis is reduced due to the application of a uniform evaluation method without considering the differences in resolution and quality of various images.

[0005] In clinical settings where faster diagnosis and treatment based on that diagnosis affect the patient's prognosis, taking FLAIR images in addition to DWI not only increases the unnecessary cost burden on the patient but also may lead to missing the opportunity for prompt early treatment.

[0006] The present invention is intended to solve the above-described problem, and provides a method for automatically classifying DFM (Diffusion-FLAIR Mismatch) for stroke diagnosis by classifying whether there is a match between a diffusion weighted image (DWI) and a FLAIR image of a stroke patient through deep learning-based artificial intelligence modeling, and a system supporting the same.

[0007] The present invention is intended to solve the above-described problem, and the purpose of the present invention is to provide a DB0M automatic classification method and a system supporting the same that quickly determines whether there is a DFM (Diffusion-FLAIR Mismatch), which is used as an important indicator for determining treatment indications for acute stroke patients, by utilizing deep learning-based artificial intelligence modeling using the mismatch between a DWI image and a B0 image (referred to as Diffusion-B0 Mismatch) without taking a FLAIR image.

[0008] The present invention is intended to solve the above-described problem, and the purpose of the present invention is to provide a rapid DAM automatic classification method and a system supporting the same, which quickly determines whether DFM (Diffusion-FLAIR Mismatch), which is used as an important indicator for determining treatment indications for acute stroke patients, is present by utilizing deep learning-based artificial intelligence modeling using the mismatch (hereinafter referred to as DAM) between a standard DWI image and an ADC map image without taking a FLAIR image.

[0009] The purpose of the present invention is not limited to the purposes mentioned above, and other purposes not mentioned will be clearly understood by those skilled in the art from the description below.

[0010] According to one embodiment of the present invention for achieving the above technical task, a DFM automatic classification system for stroke diagnosis includes an image data input unit for receiving a DWI (Diffusion-Weighted Imaging) image and a FLAIR (Fluid-Attenuated Inversion Recovery) image, a data combining unit for generating combined image data by combining the input DWI image and the FLAIR image, a DFM classification modeling unit for deriving a DFM inference value from the generated combined image data using a DFM (Diffusion-FLAIR Mismatch) classification model, and a DFM classification unit for performing DFM classification based on the derived DFM inference value.

[0011] In addition, the DFM automatic classification system for stroke diagnosis according to one embodiment may further include a preprocessing unit that preprocesses the input image data.

[0012] Additionally, the preprocessing unit can perform at least one of resizing, interpolation, normalization, and registration on the image data.

[0013] Additionally, the DFM classification model may be a pre-trained deep neural network model using a training DWI image collection and a training FLAIR image collection.

[0014] Additionally, pre-trained deep neural network models can be implemented using Convolutional Neural Networks (CNN).

[0015] In addition, the DFM classification unit can compare the derived DFM inference value with a predetermined DFM classification criterion, and if the DFM inference value is greater than the DFM classification criterion, classify the DWI image and the FLAIR image as matching, and if the DFM inference value is equal to or less than the DFM classification criterion, classify the DWI image and the FLAIR image as mismatch.

[0016] According to an embodiment of the present invention for achieving the above technical task, a DFM automatic classification method for stroke diagnosis may include a step of performing preprocessing on received DWI (Diffusion-Weighted Imaging) images and FLAIR (Fluid-Attenuated Inversion Recovery) images, a step of combining the preprocessed DWI images and FLAIR images to generate combined image data, a step of deriving a DFM inference value based on the generated combined image data and a pre-learned DFM (Diffusion-FLAIR Mismatch) classification model, and a step of performing DFM classification based on the derived DFM inference value.

[0017] In addition, the step of performing preprocessing on the received DWI image and FLAIR image may perform at least one of resizing, interpolation, normalization, and registration on the received DWI image and FLAIR image.

[0018] In addition, the step of deriving a DFM inference value based on the generated combined image data and a pre-trained DFM (Diffusion-FLAIR Mismatch) classification model may include a step of inputting the combined image data into a pre-trained deep neural network model using a training DWI image collection and a training FLAIR image collection, and a step of deriving a DFM inference value through an inference process of the deep neural network model.

[0019] Additionally, it can be implemented using the deep neural network CNN (Convolutional Neural Network).

[0020] In addition, the step of performing DFM classification based on the derived DFM inference value may include the step of comparing the derived DFM inference value with a predetermined DFM classification criterion, the step of classifying the DWI image and the FLAIR image as matching if the DFM inference value is greater than the DFM classification criterion, and the step of classifying the DWI image and the FLAIR image as mismatch if the DFM inference value is equal to or less than the DFM classification criterion.

[0021] According to another embodiment of the present invention for achieving the above technical task, a DB0M automatic classification system for rapid stroke diagnosis includes an image data input unit for receiving B0 image data and MRI diffusion weighted DWI image data, a data combining unit for generating combined image data by combining the input B0 image data and MRI diffusion weighted DWI image data, a DB0M classification modeling unit for receiving the generated combined image data and deriving a DB0M inference value using a pre-learned DB0M classification model, and a DB0M classification unit for performing DB0M classification based on the derived DB0M inference value.

[0022] In addition, the DB0M automatic classification system for stroke diagnosis according to another embodiment may further include a preprocessing unit that preprocesses the input B0 image data and MRI diffusion weighted DWI image data.

[0023] Additionally, the preprocessing unit can perform at least one of size adjustment, interpolation, normalization, and registration on the input B0 image data and MRI diffusion weighted DWI image data.

[0024] Additionally, the pre-trained DB0M classification model may be a deep neural network model pre-trained using at least one of multiple DWI images and B0 images, with the DFM set as a reference label.

[0025] Additionally, the above pre-learned deep neural network model can be implemented using a convolutional neural network (CNN).

[0026] In addition, the DB0M classification unit can compare the derived DB0M inference value with a predetermined DB0M classification criterion, and classify it as a DB0M match if the DB0M inference value is greater than the DB0M classification criterion, and classify it as a DB0M mismatch if the DB0M inference value is equal to or less than the DB0M classification criterion.

[0027] According to another embodiment of the present invention for achieving the above technical task, a DB0M automatic classification method for stroke diagnosis includes the steps of receiving B0 image data and MRI diffusion-weighted DWI image data, performing preprocessing on the input B0 image data and DWI image data, generating combined image data by combining the preprocessed B0 image data and DWI image data, deriving a DB0M inference value using the generated combined image data and a pre-learned DB0M (Diffusion-B0 Mismatch) classification model, and performing DB0M classification based on the derived DB0M inference value.

[0028] In addition, the step of performing preprocessing on the input B0 image data and DWI image data may perform at least one of size adjustment, interpolation, normalization, and registration on the input B0 image data and DWI image data.

[0029] In addition, the step of deriving a DB0M inference value using the generated combined image data and the pre-trained DB0M classification model may include the step of inputting the combined image data into a pre-trained deep neural network model using at least one of a plurality of DWI images and B0 images, and the step of deriving a DB0M inference value through the inference process of the deep neural network model.

[0030] Additionally, the above deep neural network model can be implemented using a CNN (Convolutional Neural Network).

[0031] In addition, the step of performing DB0M classification based on the derived DB0M inference value may include the step of comparing the derived DB0M inference value with a predetermined DB0M classification criterion, the step of classifying as DB0M match if the DB0M inference value is greater than the DB0M classification criterion, and the step of classifying as DB0M mismatch if the DB0M inference value is equal to or less than the DB0M classification criterion.

[0032] According to another embodiment of the present invention for achieving the above technical task, a DAM automatic classification system for rapid stroke diagnosis includes an ADC map generation unit for generating ADC map image data based on an ADC (Apparent Diffusion Coefficient) value derived by measuring signal attenuation due to diffusion from DWI (Diffusion-Weighted Imaging) image data to which different MRI (Magnetic Resonance Imaging) diffusion weights are applied, an image data input unit for receiving the generated ADC map image data and standard DWI image data to which MRI signal weights are applied related to a signal intensity of a region where water molecule diffusion is restricted in acute stroke exceeding a predetermined reference value, a data combining unit for generating combined image data by combining the input standard DWI image data and ADC map image data, a DAM classification modeling unit for receiving the generated combined image data and deriving a DAM inference value using a DAM (Diffusion-ADC Mismatch) classification model, and a DAM classification unit for performing DAM classification based on the derived DAM inference value.

[0033] In addition, the DAM automatic classification system for rapid stroke diagnosis according to another embodiment of the present invention may further include a preprocessing unit that preprocesses the input standard DWI image data and ADC map image data.

[0034] Additionally, the preprocessing unit can perform at least one of resizing, interpolation, normalization, and registration on the input standard DWI image data and ADC map image data.

[0035] Additionally, the DAM classification model may be a pre-trained deep neural network model that sets DFM as a reference label and uses multiple training DWI images and training ADC images.

[0036] Additionally, pre-trained deep neural network models can be implemented using Convolutional Neural Networks (CNN).

[0037] In addition, the DAM classification unit can compare the derived DAM inference value with a predetermined DAM classification criterion, and classify it as a DAM match if the DAM inference value is greater than the DAM classification criterion, and classify it as a DAM mismatch if the DAM inference value is equal to or less than the DAM classification criterion.

[0038] According to another embodiment of the present invention for achieving the above technical task, a DAM automatic classification method for rapid stroke diagnosis includes the steps of receiving standard DWI image data and ADC map image data to which MRI signal weights related to a signal intensity higher than a predetermined reference value in an area where water molecule diffusion is restricted in acute stroke are applied, performing preprocessing on the input standard DWI image data and ADC map image data, generating combined image data by combining the preprocessed standard DWI image data and ADC map image data, deriving a DAM inference value using the generated combined image data and a pre-learned DAM (Diffusion-ADC Mismatch) classification model, and performing DAM classification based on the derived DAM inference value.

[0039] Additionally, ADC map image data can be generated based on ADC values ​​derived by measuring signal attenuation due to diffusion from DWI image data to which different MRI diffusion weights are applied.

[0040] Additionally, the step of performing preprocessing on the input standard DWI image data and ADC map image data may perform at least one of resizing, interpolation, normalization, and registration on the input standard DWI image data and ADC map.

[0041] In addition, the step of deriving a DAM inference value using the generated combined data and the pre-trained DAM classification model may include a step of inputting the combined image data into a pre-trained deep neural network model using a plurality of training DWI images and training ADC images, and a step of deriving a DAM inference value through an inference process of the deep neural network model.

[0042] Additionally, deep neural network models can be implemented using CNN (Convolutional Neural Network).

[0043] In addition, the step of performing DAM classification based on the derived DAM inference value may include the step of comparing the derived DAM inference value with a predetermined DAM classification criterion, the step of classifying as DAM match if the DAM inference value is greater than the DAM classification criterion, and the step of classifying as DAM mismatch if the DAM inference value is equal to or less than the DAM classification criterion.

[0044] According to the present invention, the presence or absence of DFM (Diffusion-FLAIR Mismatch) between diffusion-weighted imaging (DWI) and FLAIR images of a stroke patient can be automatically classified through deep learning-based artificial intelligence modeling, thereby enabling the identification of indications for reperfusion treatment of acute stroke without being affected by differences in resolution and quality of the images to be read and the subjective judgment of each medical professional.

[0045] According to the present invention, the classification of DB0M (Diffusion-B0 Mismatch) can be quickly performed using DWI images, B0 images, and deep learning-based artificial intelligence modeling without taking FLAIR images, thereby enabling a quick determination of treatment indications for acute stroke patients.

[0046] According to the present invention, the classification of DAM (Diffusion-ADC Mismatch) can be performed using standard DWI images, ADC map images, and deep learning-based artificial intelligence modeling without taking FLAIR images, thereby enabling rapid determination of treatment indications for acute stroke patients without being affected by differences in resolution and quality of images to be read and the subjective judgment of individual medical professionals.

[0047] In addition to this, various effects may be provided that are directly or indirectly identified through this document.

[0048] FIG. 1 is a configuration diagram of a DFM automatic classification system for stroke diagnosis according to one embodiment of the present invention.

[0049] Figure 2 is a configuration diagram of a DFM automatic classification system for stroke diagnosis, which explains the data combination unit, DFM classification modeling unit, and DFM classification unit of Figure 1.

[0050] Figure 3 is an example of an input photograph for DFM judgment in stroke diagnosis according to the prior art.

[0051] Figure 4 is a flowchart of a DFM automatic classification method for stroke diagnosis according to one embodiment of the present invention.

[0052] FIG. 5 is a configuration diagram of a DB0M automatic classification system for rapid stroke diagnosis according to another embodiment of the present invention.

[0053] Figure 6 is a configuration diagram of a DBOM automatic classification system for stroke diagnosis, which explains the configuration of the data combination unit, DBOM classification modeling unit, and DBOM classification unit of Figure 5.

[0054] Figure 7 is a flowchart of a DB0M automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention.

[0055] FIG. 8 is a configuration diagram of a DAM automatic classification system for rapid stroke diagnosis according to another embodiment of the present invention.

[0056] Fig. 9 is a configuration diagram of a DAM automatic classification system for stroke diagnosis, which explains the configuration of the data combination unit, DAM classification modeling unit, and DAM classification unit of Fig. 8.

[0057] Figure 10 is a flowchart of a DAM automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention.

[0058] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. The advantages and features of the present invention, and methods for achieving them, will become clear with reference to the embodiments described in detail below together with the attached drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms. These embodiments are provided only to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Hereinafter, like reference numerals refer to like elements.

[0059] Although the terms first, second, etc. are used to describe various elements, components, and / or sections, these elements, components, and / or sections are not limited by these terms. These terms are only used to distinguish one element, component, or section from other elements, components, or sections. Accordingly, it should be understood that a first element, a first component, or a first section referred to below may also be a second element, a second component, or a second section within the technical spirit of the present invention.

[0060] The terminology used herein is for the purpose of describing embodiments only and is not intended to be limiting of the present invention. In this specification, the singular also includes the plural unless the context clearly dictates otherwise. As used herein, the terms "comprises" and / or "made of" do not exclude the presence or addition of one or more other components, steps, operations, and / or elements.

[0061] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in their common sense to those of ordinary skill in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0062] Hereinafter, the configuration of the present invention will be described in detail with reference to the attached drawings.

[0063]

[0064] FIG. 1 is a configuration diagram of a stroke diagnosis system based on DFM automatic classification according to one embodiment of the present invention.

[0065] Referring to FIG. 1, a DFM automatic classification system (100) for stroke diagnosis according to one embodiment of the present invention (hereinafter referred to as a DFM automatic classification system) includes an image data input unit (110), a data preprocessing unit (120), a data combining unit (130), a DFM classification modeling unit (140), and a DFM classification unit (150). In addition, the DFM automatic classification system (100) according to one embodiment may include a DFM classification model learning unit (160).

[0066] According to one embodiment, the image data input unit (110) of the DFM automatic classification system (100) receives (also referred to as receives) image data corresponding to two-dimensional or three-dimensional brain image data. The brain image data received by the image data input unit (110) according to one embodiment includes a DWI image and a FLAIR image (in the following embodiments, the image is merely intended to emphasize the image itself, and the image data is merely intended to emphasize the data aspect for computer processing of the corresponding image, and in the following embodiments, the image and the image data are used interchangeably and are not used as different components of the invention).

[0067] In the following examples, brain imaging data refers to image data that visualizes the internal structure and / or function of the brain through direct or indirect methods, and may include magnetic resonance imaging, computed tomography (CT) imaging, positron emission tomography (PET) imaging, single photon emission computed tomography (SPECT) imaging, etc. Magnetic resonance imaging (MRI) refers to a diagnostic technology that uses a magnetic field to create images or pictures of internal structures, and images obtained through the same.

[0068] In the following examples, the DWI image is image data obtained through a technique that measures the irregular movement of water molecules within a tissue, i.e., the diffusion velocity, and is generally an image processing method that can early measure cerebral infarction or lesions caused by cerebral vascular occlusion. By detecting such signal changes, various ischemic brain tissue damages such as the area of ​​brain tissue damage and the volume of damage can be imaged. At this time, the diffusion velocity is determined by the b value, which represents the strength of the gradient magnetic field (the b value is a parameter that controls the signal attenuation due to the diffusion of water molecules in the DWI image, and the higher the value, the greater the signal attenuation due to diffusion). In general, DWI images are known to be useful for quickly detecting acute or subacute cerebral infarction.

[0069] DWI (Diffusion-Weighted Imaging) is a technique for taking diffusion-weighted images, and can use various b-values. Typically, the b-value is 1000 s / mm. 2 (This is called B1000) is known to be appropriate for emphasizing the area where the diffusion of water molecules is restricted, and for this reason, it is a b-value widely used in stroke diagnosis (however, various b-values ​​can be applied for diffusion weighting when acquiring DWI images so that the movement of water molecules in the tissue can be visualized from various angles by weighting diffusion differently, and various b-values ​​such as B500, B1000, and B2000 can be used depending on the characteristics of the disease, diagnostic purpose, and specifications of the MRI equipment).

[0070] The DWI image input into the image data input unit (110) of the DFM automatic classification system (100) according to one embodiment may be a DWI image to which MRI signal weighting related to a signal intensity higher than a predetermined standard value is applied in an area where water molecule diffusion is restricted in acute stroke (generally, the predetermined standard value is a b value of 1000 s / mm 2(This may be a B1000 image, and such thresholds may vary depending on implementation).

[0071] In the examples below, the FLAIR image is one of the MRI techniques, which attenuates the signal of a body fluid such as cerebrospinal fluid (CSF) by a set value of the inversion time (TI), which is a variation of the inversion recovery sequence, to make the CSF appear dark in the captured image, thereby making the difference with the contrasting lesion or tissue more prominent (the inversion time refers to the time required to reduce the signal of a specific tissue in an MRI scan as much as possible, and by adjusting this time when acquiring a FLAIR image, the signal of the CSF can be eliminated, and as a result, pathological changes in the brain can be more clearly emphasized).

[0072] This attenuation of CSF signals allows for more clearly highlighting brain lesions or tissue changes, and enhances the contrast between white and gray matter, as well as between lesions and normal tissue, allowing for a more accurate assessment of various brain pathological conditions. FLAIR imaging is generally known to be useful for assessing the temporal course of cerebral infarction.

[0073] The data preprocessing unit (120) of the DFM automatic classification system (100) according to one embodiment can perform preprocessing on brain image data, such as DWI images and FLAIR images, received from the image data input unit (110) so that they can be used in the DFM classification modeling unit (140), which is an artificial intelligence modeling unit to be described later.

[0074] The preprocessing performed by the data preprocessing unit (120) according to one embodiment may perform at least one of image size adjustment, interpolation, normalization, and image registration.

[0075] That is, the data preprocessing unit (120) according to one embodiment can adjust the size of each image to a predetermined resolution (e.g., 256 Х 256) and then perform interpolation to configure each image data (DWI image, FLAIR image) into a plurality of images (e.g., 36) when the resolutions of the DWI image and FLAIR image of the input image data are not uniform (resize) (at this time, the nearest neighborhood method is used in the resize and interpolation process to enlarge or reduce the resolution and depth of the data format). After this, each image can be restricted to have a value between 0 and 1 through image-by-image normalization for all data formats.

[0076] In general, the shape of the brain expressed in a FLAIR image is different from that of a DWI image, and since this difference may hinder the artificial intelligence model from achieving optimal performance, in order to minimize the influence due to the difference in image characteristics, the data preprocessing unit (120) according to one embodiment may perform image registration as a matching method to match the shape of the brain in the DWI image with the shape of the brain in the FLAIR image. Image registration is a process of aligning corresponding points between two or more images, establishing a geometric relationship between the acquired images, and aligning them into a common coordinate system.

[0077] In addition, the data preprocessing unit (120) according to one embodiment may perform denoising, intensity normalization, skull stripping, bias field correction, feature extraction, data augmentation, etc. (however, it is not necessarily limited thereto, and any preprocessing process necessary for applying an artificial intelligence model may be included and performed).

[0078] Noise removal, which can be performed as a preprocessing step, is a process of reducing noise in image data and improving the quality of image data by performing noise removal techniques such as Gaussian filter and non-local mean filter to remove noise that may be included in the process of acquiring image data.

[0079] Intensity normalization, which can be performed as a preprocessing step, is the process of standardizing image data intensity to reduce intensity differences between image data acquired at different times or from different image acquisition devices. Skull stripping, which can be performed as a preprocessing step, removes non-brain tissue, such as the skull and skin, from brain images, making it necessary when analyzing only brain tissue. Bias field correction, which can be performed as a preprocessing step, improves image uniformity by adjusting non-uniform signal intensities during image data acquisition to minimize the influence of bias fields.

[0080] Feature extraction, which can be performed as a preprocessing step, is the process of extracting useful features (e.g., information about edges, texture information, etc.) from an image before inputting it into a deep learning or machine learning model. These feature extractions can be utilized for analysis such as lesion classification and segmentation. Data augmentation, which can be performed as a preprocessing step, is the process of generating more training data by transforming existing image data through methods such as rotation, inversion, and resizing to improve the generalization ability of an artificial intelligence model.

[0081] The data combining unit (130) of the DFM automatic classification system (100) according to one embodiment can combine the DWI image (B1000 with diffusion weighting reflected) and the FLAIR image preprocessed in the data preprocessing unit (120) to generate combined image data, as shown in FIG. 2.

[0082] That is, the data combining unit (130) according to one embodiment can combine (merge) a DWI image (for example, an image with dimensions of (1, 36, 256, 256)) and a FLAIR image (for example, an image with dimensions of (1, 36, 256, 256)) that have been resized and interpolated to match the same dimension to generate combined image data (for example, an image with dimensions of (2, 36, 256, 256)) (the adjusted size, the interpolation method, etc. can be implemented in various ways depending on the implementation level). Meanwhile, the order of data preprocessing and data combining is not necessarily limited to performing preprocessing and then combining data, and can be implemented so that preprocessing is performed on the combined image data after combining the DWI image and the FLAIR image, depending on the implementation method.

[0083] The DFM classification modeling unit (140) of the DFM automatic classification system (100) according to one embodiment performs inference to derive a DFM inference value using a pre-learned DFM classification model for combined image data input from the data combining unit (130).

[0084] As described above, the consistency between DWI images and FLAIR images is used as an important indicator for determining indications for acute stroke treatment. The DFM classification modeling unit (140) according to one embodiment can derive a DFM inference value indicating the consistency between the DWI image and the FLAIR image by using a DFM (Diffusion-Flair Mismatch) classification model that classifies the consistency between the DWI image and the FLAIR image. At this time, the DFM classification model may be a deep neural network model that is pre-trained using a training DWI image collection and a training FLAIR image collection, as described below.

[0085] In one embodiment, a DFM classification model used in the DFM classification modeling unit (140) may be a convolutional neural network (CNN), and a deep learning-based 3D convolutional neural network (CNN) such as 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception may be used. Such a 3D CNN model can process 3D data to recognize and learn complex patterns, to learn and recognize complex structures in medical image analysis, to be useful for capturing important structural features in high-resolution data such as medical images, and to extract features of images and solve classification or regression problems using multiple convolutional layers, and to improve the training process by optimizing the way information flows throughout the depth of the network (for example, DenseNet directly transfers information between all layers, and ResNet uses a residual connection method that adds the input to the output).

[0086] The DFM classification modeling unit (140) according to one embodiment may use a 3D-DenseNet model as illustrated in FIG. 2. That is, the DFM classification modeling unit according to one embodiment uses a 3D DenseNet as a binary classification model for DFM, and at this time, the CNN network architecture may be composed of a 4-repeated dense block that uses the connected data of DWI and FLAIR as input and a transition layer.

[0087] A DFM classification modeling unit (140) according to one embodiment may have a basic structure such as a dense block composed of two convolutional layers (i.e., 1×1×1 and 3×3×3 convolutional layers), and this structure may be implemented so that the depth of the block increases by repeating. At this time, a plurality of dense blocks (for example, 4) of the CNN network may have various depths depending on the order, and may have two convolutional layers that are repeated 6, 12, 24, and 16 times, respectively, depending on the order of the blocks (the number of repetitions may be changed depending on the implementation method), and the last linear layer may have a plurality of nodes (for example, 2) and may be implemented so as to classify matching and mismatching cases depending on the scale of the value of each node. At this time, in the DFM classification modeling unit (140) according to one embodiment, batch normalization (BN) and 3D-Conv (3-dimensional convolution) can be applied, and a stride of 2 (but not necessarily limited thereto) can be uniformly used in the convolution layer and pooling layer excluding the dense block and transition layers.

[0088] In general, the final output layer of the neural network can calculate the probability for each class (matched and unmatched) through a linear layer and finally using a softmax activation function, and finally perform classification (also called discrimination) based on the calculated probability value. In this way, the DFM classification modeling unit (140) according to one embodiment inputs the input image combination data into a 3D CNN model, and as shown in FIG. 2, it inputs the DFM inference value ( ) is derived.

[0089] The DFM classification unit (150) of the DFM automatic classification system (100) according to one embodiment performs DFM classification based on the DFM inference value transmitted from the DFM classification modeling unit (140). That is, the DFM classification unit (150) according to one embodiment performs DFM classification based on the transmitted DFM inference value (as shown in FIG. 2). ) is classified as a matched case if it is greater than a given classification criterion (e.g. 0.7). ), if it is below the classification criterion, it is classified as a mismatch (unmatched case). ) can be done (depending on the implementation method, the classification method can be set differently for cases that match and do not match the classification criteria)

[0090] In general, if the signal intensity according to the lesion location in the DWI image matches the signal intensity according to the lesion location in the FLAIR image, it can be classified as a match, and if the DWI image is a bright image (corresponding to high signal intensity) but the FLAIR image is not bright (iso signal), it can be classified as a mismatch. That is, as shown in the example of (a) DFM match in Fig. 3, if the signal intensity according to the lesion location in the FLAIR image and the signal intensity location according to the lesion location in the DWI image are significantly similar, the DFM classification result can be classified as a match, and as shown in the example of (b) DFM mismatch in Fig. 3, if the signal intensity according to the lesion location in the FLAIR image and the signal intensity location according to the lesion location in the DWI image are significantly similar, the DFM classification result can be classified as a mismatch.

[0091] The DFM automatic classification system (100) according to one embodiment may further include a DFM classification model learning unit (160). The DFM classification model learning unit (160) according to one embodiment trains the DFM classification model using learning data. That is, the DFM classification model learning unit (160) according to one embodiment may train the DFM classification model using a plurality of learning DWI image collections and learning FLAIR image collections. That is, in a medical field, at least one of learning FLAIR images, learning DWI image data, and learning B0 images for a plurality of clinical patients is reviewed to determine the presence or absence of DFM, and the resulting generated ground truth can be used as an accurate label of the DFM and as a training dataset for the DFM classification model of the present invention.

[0092] In addition, the DFM classification model learning unit (160) according to one embodiment can pre-train (train) the DFM classification model by creating a pre-trained model using a method that precedes pretext training, such as self-supervised learning using a collection of natural images or MRI images such as brain DWI images and FLAIR images without lesions, and then fine-tuning the model.

[0093] In addition, the DFM classification model learning unit (160) according to one embodiment can perform processes related to de-identification of image data and data protection to safely manage medical image data and protect the patient's personal information.

[0094] In the image data de-identification process performed in the DFM classification model learning unit (160) according to one embodiment, a collected data screening process for selecting and reviewing available medical image data (Screening for collected data), a personal information removal process for downloading image files of the selected data and removing the patient's personally identifiable information from the image files to protect the privacy of the individual (Remove personal information from image files), and a labeling and subject conversion process for assigning a new label to the de-identified data or converting it into a new subject for patient identification can be performed.

[0095] The de-identified patient basic information processing process performed in the DFM classification model learning unit (160) according to one embodiment may perform a de-identified patient basic information processing process that removes personally identifiable elements from the patient's basic information to make it impossible to identify an individual, which is de-identified personal information, and anonymized information, which is information processed so that an individual can no longer be identified, and a risk of re-identification process that considers the risk of patient information being re-identified even during the de-identification process.

[0096] In addition, the DFM classification model learning unit (160) according to one embodiment may perform an operator management process for managing operators who process data, a user registration process for registering users who will use the data processing system, an inspector / operator assignment process for allocating personnel to supervise and operate each data processing process, a data processing process for processing de-identified data, annotation process for annotating data or adding metadata, and a first inspection and second inspection sampling process for checking whether the data processing and de-identification processes have been performed appropriately.

[0097] In this way, the DFM classification model learning unit (160) according to one embodiment can perform various processes for protecting personal information and ensuring data security as described above, and, in contrast, a separate configuration module (not shown) can perform various processes for protecting personal information and ensuring data security for data used in the learning and inference process.

[0098] Meanwhile, the DFM classification model learning unit (160) according to one embodiment can automatically or manually label data based on the characteristics of different tissues or lesions exhibiting different signal intensities in medical images such as MRI or CT. By doing so, specific tissues or lesions can be classified and information necessary for medical diagnosis can be extracted (in this case, signal intensity discrepancies refer to differences compared to a specific standard or expected value, and data points can be assigned appropriate labels based on these differences).

[0099] According to one embodiment, the DFM classification model learning unit (160) may use a cross-validation method such as k-fold cross-validation. In this case, the data set for learning is divided into multiple partitions (folds), and each partition may be used for training, tuning, and testing different parts of the data set (Training Set), thereby increasing the generalization ability of the DFM classification model by allowing the DFM classification model to learn and be tested on various parts of the data.

[0100] According to one embodiment, a DFM classification model learning unit (160) can learn by dividing a dataset including a plurality of learners (stroke patients) into N segments (folds) and using a dataset configured such that each segment functions as a training set (N-1) and a tuning and test set (1). The training set is a dataset used to learn an artificial intelligence model, the tuning set is a dataset used to adjust hyperparameters of the artificial intelligence model, and the test set is a dataset used to evaluate how well the artificial intelligence model generalizes.

[0101] Specifically, when splitting the dataset into training, training, and test sets, a similar proportion of labels can be maintained in each set to maintain distributional consistency, and the predicted probabilities on the test set of the DFM classifier derived from each split are aggregated to form a cohort consisting of a large number of unique trainees, and performance metrics are applied to this aggregated data to provide an overall evaluation of the model's predictive ability. This comprehensive approach can emphasize the robustness of the performance evaluation by reflecting the model's generalizability to new, unseen data.

[0102] Meanwhile, the DFM classification model learning unit (160) according to one embodiment may use other data set configuration methods that can be used for learning an artificial intelligence model in addition to the above-described method, such as a simple holdout method, a stratified k-fold cross-validation, a leave-p-out cross-validation, a time-based split, bootstrap, and Monte Carlo cross-validation (however, the present invention is not necessarily limited thereto, and other data set configuration methods for learning an artificial intelligence model that are not mentioned may be used). The DFM classification model learning unit (160) according to one embodiment may appropriately select a data set configuration method according to the characteristics of the data, the complexity of the model, the precision requirements of the evaluation, and the like so that the model can work well on real-world data.

[0103] Hereinafter, a DFM automatic classification method for stroke diagnosis according to one embodiment of the present invention is described based on the above-described description.

[0104]

[0105] FIG. 4 is an example of a DFM automatic classification method for stroke diagnosis according to an embodiment of the present invention, which is an example of a DFM automatic classification method for stroke diagnosis using a DFM automatic classification device for stroke diagnosis according to an embodiment of the present invention of FIG. 1.

[0106] Referring to FIG. 4, a DFM automatic classification method for stroke diagnosis according to an embodiment of the present invention includes a step of performing preprocessing on received DWI images and FLAIR images, a step of combining the preprocessed DWI images and FLAIR images to generate combined image data, a step of deriving a DFM inference value based on the generated combined image data and a pre-learned DFM classification model, and a step of performing DFM classification based on the derived DFM inference value.

[0107] Below, each step of the DFM automatic classification method for stroke diagnosis of Fig. 4 will be described in detail. First, DWI images and FLAIR images are received by the image data input unit (110) (S410).

[0108] Preprocessing is performed on the DWI image and FLAIR image received by the data preprocessing unit (120) (S420). When preprocessing the DWI image and FLAIR image by the data preprocessing unit (120), at least one of size adjustment, interpolation, normalization, and registration may be performed on the received DWI image and FLAIR image.

[0109] The DWI image and FLAIR image preprocessed by the data combining unit (130) are combined to generate combined image data (S430). The DFM classification modeling unit (140) derives a DFM inference value based on the combined image data and the pre-learned DFM classification model (S440).

[0110] At this time, when deriving a DFM inference value using a DFM classification model pre-trained by the DFM classification modeling unit (140), the training DWI image and the training FLAIR image (each and / or when combined) are input into an implemented deep neural network model using at least one of pre-trained CNN, 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception, and the DFM inference value can be derived through the inference process of the corresponding deep neural network model.

[0111] Thereafter, DFM classification is performed based on the DFM inference value derived by the DFM classification unit (150) (S450). For example, when performing DFM classification based on the derived DFM inference value, the derived DFM inference value is compared with a predetermined DFM classification criterion, and if the resulting DFM inference value is greater than the DFM classification criterion, it is classified as a match between the DWI image and the FLAIR image, and if the DFM inference value is equal to or less than the DFM classification criterion, it is classified as a mismatch between the DWI image and the FLAIR image.

[0112] The present embodiment described above can automatically classify whether there is a DFM (Diffusion-FLAIR Mismatch) between a diffusion-weighted image (DWI) and a FLAIR image of a stroke patient through deep learning-based artificial intelligence modeling, thereby identifying indications for reperfusion treatment of acute stroke without being affected by differences in resolution and quality of the images to be read and the subjective judgment of each medical professional.

[0113] FIG. 5 is a configuration diagram of a stroke diagnosis system based on DB0M automatic classification according to another embodiment of the present invention.

[0114] Referring to FIG. 5, a DB0M automatic classification system (500) for rapid stroke diagnosis according to another embodiment of the present invention (hereinafter referred to as the DB0M automatic classification system) includes an image data input unit (510), a data preprocessing unit (520), a data combining unit (530), a DB0M classification modeling unit (540), and a DB0M classification unit (550). In addition, the DB0M automatic classification system (500) according to one embodiment may include a DB0M classification model learning unit (560).

[0115] MRI refers to a diagnostic technology that uses a magnetic field to create an image or photograph of a structure in the body, and the image obtained thereby. The image data input unit (110) of the DB0M automatic classification system (500) according to another embodiment receives MRI (Magnetic Resonance Imaging) non-diffusion weighted B0 image data and MRI diffusion weighted DWI image data (in the following embodiments, the image emphasizes the image itself, and the image data emphasizes the data aspect for computer processing of the image. In the following embodiments, the image and the image data are used interchangeably and are not used as different components of the invention).

[0116] B0 images can be used as a substitute for T2-weighted images. T2-weighted images are images used to visualize the properties of materials in MRI. In T2 images, tissues rich in water, such as water, appear bright (high signal intensity) because water molecules lose energy slowly (long T2 relaxation time), while tissues in which water molecules lose energy quickly appear dark (low signal intensity). B0 image data can be acquired as part of a DWI image data sequence, where B0 means no diffusion weighting (meaning 'zero B-value'). B0 image data is used as a reference image for a DWI image data sequence, and reflects the original signal intensity of the tissue without signal changes based on the degree of diffusion of water molecules.

[0117] B0 images, similar to T2-weighted images, can highlight water-rich tissues and emphasize water signals, and provide high contrast like T2-weighted images, which can help distinguish lesions or abnormal tissues. FLAIR is basically a T2-weighted image in which the cerebrospinal fluid signal is nulled. Due to these characteristics, B0 image data can be used as a substitute for T2-weighted images for rapid DWI classification in certain situations, such as acute stroke, by utilizing B0 image data obtained from a sequence of DWI images without additional T2-weighted images acquisition (since additional shooting is not required, MRI scan time can be shortened and resources can be saved).

[0118] In other embodiments below, B0 image data and MRI diffusion weighted DWI image data are input, but are not necessarily limited to the MRI method, and image data visualizing the internal structure and / or function of the brain can be input through direct or indirect methods such as computed tomography (CT) images, positron emission tomography (PET) images, and single photon emission computed tomography (SPECT) images.

[0119] In other embodiments below, the DWI image is image data obtained through a technique that measures the irregular movement of water molecules in a tissue, i.e., the diffusion velocity, and is generally an image processing method that can early measure cerebral infarction or lesions caused by cerebral vascular occlusion. By detecting such signal changes, various ischemic brain tissue damages such as the brain tissue damage area and damage volume can be imaged. At this time, the diffusion velocity is determined by the B value indicating the strength of the gradient magnetic field (the B value is a parameter that controls signal attenuation due to the diffusion of water molecules in the DWI image, and the higher the value, the greater the signal attenuation due to diffusion). In general, DWI images are known to be useful for quickly detecting acute or subacute cerebral infarction.

[0120] DWI (Diffusion-Weighted Imaging) is a technique for taking diffusion-weighted images, and various B values ​​can be used. Generally, a B value of 1000 s / mm² (referred to as B1000) is known to be appropriate for emphasizing the area where the diffusion of water molecules is restricted, and for this reason, it is a B value that is widely used in stroke diagnosis (however, various B values ​​can be applied for diffusion weighting when acquiring DWI images so that the movement of water molecules in the tissue can be visualized from various angles by weighting diffusion differently, and various B values ​​such as B500, B1000, and B2000 can be used depending on the characteristics of the disease, diagnostic purpose, and specifications of the MRI equipment.)

[0121] According to another embodiment, a DWI image input into the image data input unit (510) of the DB0M automatic classification system (500) may be a DWI image to which MRI signal weighting is applied in relation to a signal intensity higher than a predetermined reference value in an area where water molecule diffusion is restricted in acute stroke (generally, the predetermined reference value may be a B1000 image with a B value of 1000 s / mm, and such a reference value may be changed depending on the implementation). In the embodiments below, the DWI image means a diffusion-weighted DWI image to which MRI diffusion signal weighting is applied in relation to a signal intensity higher than a predetermined reference value.

[0122] The data preprocessing unit (520) of the DB0M automatic classification system (500) according to another embodiment can perform a preprocessing process so that the B0 image data and DWI image data received from the image data input unit (510) can be used in the DB0M classification modeling unit (540), which is an artificial intelligence modeling unit to be described later.

[0123] The preprocessing process performed by the data preprocessing unit (520) according to another embodiment may perform at least one of image size adjustment, interpolation, normalization, and image registration.

[0124] That is, the data preprocessing unit (520) according to another embodiment can adjust the size of each image (e.g., 256 × 256) when the resolution of the input B0 image data and DWI image data is not uniform, and then perform interpolation to configure each image data (B0 image data and DWI image data) into a plurality of images (e.g., 36) (at this time, the nearest neighborhood method is used in the resizing and interpolation process to enlarge or reduce the resolution and depth of the data format). After this, each image can be restricted to have a value between 0 and 1 through image-by-image normalization for all data formats.

[0125] In addition, the data preprocessing unit (520) according to another embodiment may perform denoising, intensity normalization, skull stripping, bias field correction, feature extraction, data augmentation, etc. (however, it is not necessarily limited thereto, and any preprocessing process necessary for applying an artificial intelligence model may be included and performed).

[0126] Noise removal, which can be performed as a preprocessing step, is a process of reducing noise in image data and improving the quality of image data by performing noise removal techniques such as Gaussian filter and non-local mean filter to remove noise that may be included in the process of acquiring image data.

[0127] Intensity normalization, which can be performed as a preprocessing step, is the process of standardizing image data intensity to reduce intensity differences between image data acquired at different times or from different image acquisition devices. Skull stripping, which can be performed as a preprocessing step, removes non-brain tissue, such as the skull and skin, from brain images, making it necessary when analyzing only brain tissue. Bias field correction, which can be performed as a preprocessing step, improves image uniformity by adjusting non-uniform signal intensities during image data acquisition to minimize the influence of bias fields.

[0128] Feature extraction, which can be performed as a preprocessing step, is the process of extracting useful features (e.g., information about edges, texture information, etc.) from an image before inputting the image to a deep learning or machine learning model. Such feature extraction can be utilized for analysis such as lesion classification and segmentation. Data augmentation, which can be performed as a preprocessing step, is the process of generating more training data by transforming existing image data through methods such as rotation, inversion, and resizing to improve the generalization ability of an artificial intelligence model.

[0129] The data combining unit (530) of the DB0M automatic classification system (500) according to another embodiment can combine B0 image data and DWI image data preprocessed in the data preprocessing unit (520) to generate combined image data.

[0130] That is, the data combining unit (530) according to another embodiment can combine (merge) B0 image data (for example, dimensions of (1, 36, 256, 256)) and DWI image data (for example, dimensions of (1, 36, 256, 256)) that have been resized and interpolated to match the same dimension to generate combined image data (for example, combined image data of dimensions of (2, 36, 256, 256)) (the adjusted size and interpolation method, etc., can be implemented in various ways depending on the implementation level).

[0131] Meanwhile, the order of data preprocessing and combining is not necessarily limited to performing preprocessing and then combining data, and depending on the implementation method, preprocessing may be performed on combined image data generated after combining DWI image data and B0 image data.

[0132] According to another embodiment, the DB0M classification modeling unit (540) of the DB0M automatic classification system (500) performs inference to derive a DB0M inference value by modeling (inferring) the combined image data input from the data combining unit (530) using a pre-learned DB0M classification model.

[0133] As mentioned above, the consistency between DWI and FLAIR images is used as an important indicator for determining the timing of stroke treatment. The DB0M classification modeling unit (540) according to one embodiment can derive an inference value indicating whether or not DB0M is present by using a pre-trained model that classifies DB0M using the DWI and B0 images as input.

[0134] The DB0M classification model is a deep neural network model that is pre-trained using multiple training DWI images and training B0 images, with DFM set as the reference label.

[0135] In another embodiment, a DB0M classification model used in the DB0M classification modeling unit (540) may be a convolutional neural network (CNN), and a deep learning-based 3D convolutional neural network (CNN) such as 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception may be used. Such a 3D CNN model can process 3D data to recognize and learn complex patterns, to learn and recognize complex structures in medical image analysis, to be useful for capturing important structural features in high-resolution data such as medical images, and to extract features of images and solve classification or regression problems using multiple convolutional layers, and to improve the training process by optimizing the way information flows throughout the depth of the network (for example, DenseNet directly passes information between all layers, and ResNet uses a residual connection method that adds the input to the output).

[0136] The DB0M classification modeling unit (540) according to another embodiment may use a 3D-DenseNet model as illustrated in FIG. 6. That is, as illustrated in FIG. 6, the DB0M classification modeling unit according to another embodiment uses 3D DenseNet as a binary classification model for DB0M, and at this time, the CNN network architecture may be composed of a 4-time repeated dense block and a transition layer that uses combined image data, which is a combination of DWI image data and B0 image data, as input.

[0137] According to another embodiment, the DB0M classification modeling unit (540) may have a basic structure such as a dense block composed of two convolutional layers (i.e., 1×1×1 and 3×3×3 convolutional layers), and this structure may be implemented so that the depth of the block increases by repeating. At this time, multiple (for example, 4) dense blocks of the CNN network may have various depths depending on the order, and may have two convolutional layers that are repeated 6, 12, 24, and 16 times respectively depending on the order of the blocks (the number of repetitions may be changed depending on the implementation), and the last linear layer may have multiple (for example, 2) nodes, and may be implemented so as to classify matching and mismatching cases depending on the scale of the value of each node.

[0138] At this time, in the DB0M classification modeling unit (540) according to another embodiment, batch normalization (BN) and 3D-Conv (3D convolution) can be applied, and a stride of 2 (but not necessarily limited thereto) can be uniformly used in the convolution layer and pooling layer excluding the dense block and transition layer.

[0139] In general, the final output layer of the neural network can calculate the probability for each class (matched and unmatched) through a linear layer and finally using a softmax activation function, and finally perform classification (also called discrimination) based on the calculated probability value. In this way, the DB0M classification modeling unit (540) according to another embodiment inputs the input combined image data into a 3D CNN model, and as shown in FIG. 6, it inputs the DB0M inference value corresponding to the probability for each class (matched and unmatched) through a linear layer and finally using a softmax activation function. ) is derived.

[0140] The DB0M classification unit (550) of the DB0M automatic classification system (500) according to another embodiment performs DB0M classification based on the DB0M inference value transmitted from the DB0M classification modeling unit (540). That is, the DB0M classification unit (550) according to another embodiment performs DB0M classification based on the transmitted DB0M inference value (as shown in FIG. 6). ) is classified as a matched case if it is greater than a given classification criterion (e.g. 0.7). ), if it is below the classification criterion, it is classified as a mismatch (unmatched case). ) can be done (depending on the implementation method, the classification method can be set differently for cases that match and do not match the classification criteria).

[0141] In this way, the DB0M classification unit (550) according to another embodiment can classify DB0M using only B0 images and DWI images without FLAIR images, based on patterns obtained in the learning process described later.

[0142] A DB0M automatic classification system (500) according to another embodiment may further include a DB0M classification model learning unit (560). The DB0M classification model learning unit (560) according to another embodiment may train a DB0M classification model using learning data (at least one of a plurality of learning DWI images, learning FLAIR images, and learning B0 images).

[0143] In addition, the DB0M classification model learning unit (560) according to another embodiment can pre-train the DB0M classification model by creating a pre-trained model using a method that precedes pretext training, such as self-supervised learning using a collection of natural images, brain DWI images without lesions, and MRI images such as B0, and then fine-tuning the model.

[0144] FLAIR is one of the MRI techniques. It attenuates the signal of body fluids such as cerebrospinal fluid (CSF) by a set value of inversion time (TI), which is a variation of the inversion recovery sequence, to make CSF appear darker in the captured image, thereby making the difference with contrasting lesions or tissues more prominent (inversion time refers to the time required to minimize the signal of a specific tissue in an MRI scan, and by adjusting this time when acquiring a FLAIR image, the signal of CSF can be eliminated, resulting in more distinctly emphasized pathological changes in the brain).

[0145] This attenuation of CSF signals allows for more clearly highlighting brain lesions or tissue changes, and enhances contrast between white and gray matter, as well as between lesions and normal tissue, allowing for a more accurate assessment of various brain pathological conditions. FLAIR and T2-weighted images are generally known to be more useful than other MR sequences for assessing the temporal course of cerebral infarction.

[0146] A method for learning and generating a DB0M classification model in a DB0M classification model learning unit (560) according to another embodiment is as follows.

[0147] Multiple medical professionals evaluate a patient's DWI and B0 images to determine the presence or absence of DB0M, and define the results as reference labels. These labels are used as ground truth to train and validate the DB0M classification model of the present invention.

[0148] A DB0M classification model learning unit (560) according to another embodiment can generate a DB0M classification model capable of recognizing a pattern between DB0M labels associated with at least one of DWI image data, FLAIR image data, B0 image data, and ADC image data (including ADC map image data) using the aforementioned training dataset. That is, a DB0M classification model learning unit (560) according to another embodiment can learn the characteristics of B0 image data and DWI image data using a machine learning or deep learning algorithm to build a model capable of classifying whether or not it is DB0M even without FLAIR image data.

[0149] The DB0M classification model learning unit (560) according to another embodiment can perform processes related to de-identification of image data and data protection to safely manage medical image data and protect the patient's personal information.

[0150] In the image data de-identification process performed in the DB0M classification model learning unit (560) according to another embodiment, a collected data screening process for selecting and reviewing available medical image data (Screening for collected data), a personal information removal process for downloading image files of the selected data and removing the patient's personal identification information from the image files to protect the privacy of the individual (Remove personal information from image files), and a labeling and subject conversion process for assigning a new label to the de-identified data or converting it into a new subject for patient identification can be performed.

[0151] The de-identified patient basic information processing process performed in the DB0M classification model learning unit (560) according to another embodiment may perform a de-identified patient basic information processing process that removes personally identifiable elements from the patient's basic information to make it impossible to identify an individual, which is de-identified personal information, and anonymized information, which is information processed so that an individual can no longer be identified, and a risk of re-identification process that considers the risk of patient information being re-identified even during the de-identification process.

[0152] In addition, the DB0M classification model learning unit (560) according to another embodiment may perform an operator management process for managing operators who process data, a user registration process for registering users who will use the data processing system, an inspector / operator assignment process for allocating personnel to supervise and operate each data processing process, a data processing process for processing de-identified data, annotation process for annotating data or adding metadata, and a first inspection and second inspection sampling process for checking whether the data processing and de-identification processes have been performed appropriately.

[0153] In this way, the DB0M classification model learning unit (560) according to another embodiment can perform various processes for protecting personal information and ensuring data security as described above, and, in contrast, a separate configuration module (not shown) can perform various processes for protecting personal information and ensuring data security for data used in the learning and inference process.

[0154] Meanwhile, the DB0M classification model learning unit (560) according to another embodiment can automatically or manually label data based on the characteristics of different tissues or lesions exhibiting different signal intensities in medical images such as MRI or CT. By doing so, specific tissues or lesions can be classified and information necessary for medical diagnosis can be extracted (in this case, signal intensity discrepancies refer to differences compared to a specific standard or expected value, and data points can be assigned appropriate labels based on these differences).

[0155] According to another embodiment, the DB0M classification model learning unit (560) may use a cross-validation method such as k-fold cross-validation. In this case, the data set for learning is divided into multiple partitions (folds), and each partition may be used for training, tuning, and testing different parts of the data set (Training Set), thereby increasing the generalization ability of the DB0M classification model by allowing the DB0M classification model to learn and be tested on various parts of the data).

[0156] According to another embodiment, a DB0M classification model learning unit (560) can learn by dividing a dataset including a plurality of learners (stroke patients) into N segments (folds) and using a dataset configured such that each segment functions as a training set (N-1) and a tuning and test set (1). The training set is a dataset used to learn an artificial intelligence model, the tuning set is a dataset used to adjust hyperparameters of the artificial intelligence model, and the test set is a dataset used to evaluate how well the artificial intelligence model generalizes.

[0157] Specifically, when splitting the dataset into training, training, and test sets, a similar proportion of labels can be maintained in each set to maintain distributional consistency, and the predicted probabilities on the test set of the DB0M classification model derived from each split are aggregated to form a cohort consisting of a large number of unique trainees, and performance metrics are applied to this aggregated data to provide an overall evaluation of the model's predictive ability. This comprehensive approach can emphasize the robustness of the performance evaluation by reflecting the model's generalizability to new, unseen data.

[0158] Meanwhile, the DB0M classification model learning unit (560) according to another embodiment may use other dataset configuration methods that can be used for learning an artificial intelligence model in addition to the aforementioned method, such as a simple holdout method, a stratified k-fold cross-validation, a leave-p-out cross-validation, a time-based split, bootstrap, and Monte Carlo cross-validation (however, the present invention is not necessarily limited thereto, and dataset configuration methods for learning an artificial intelligence model that are not mentioned may be used). The DB0M classification model learning unit (560) according to one embodiment may appropriately select a dataset configuration method according to the characteristics of the data, the complexity of the model, the precision requirements of the evaluation, and the like so that the model can work well on real-world data.

[0159] Hereinafter, a DB0M automatic classification method for stroke diagnosis according to another embodiment of the present invention is described based on the above-described description.

[0160]

[0161] FIG. 7 is an example of a DB0M automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention, which is an example of a DB0M automatic classification method for rapid stroke diagnosis using a DB0M automatic classification device for rapid stroke diagnosis according to an embodiment of the present invention of FIG. 5.

[0162] Referring to FIG. 7, a DB0M automatic classification method for rapid stroke diagnosis according to an embodiment of the present invention includes a step of receiving DWI image data and B0 image data, a step of performing preprocessing on the received DWI image data and B0 image data, a step of combining the preprocessed DWI image data and B0 image data to generate combined image data, a step of deriving a DB0M inference value using the generated combined image data and a pre-learned DB0M classification model, and a step of performing DB0M classification based on the derived DB0M inference value.

[0163] Hereinafter, each step of the DB0M automatic classification method for rapid stroke diagnosis of Fig. 7 will be described in detail. First, DWI image data and B0 image data are input (received) by the image data input unit (510) (S710).

[0164] Preprocessing is performed on the DWI image data and B0 image data received by the data preprocessing unit (520) (S720). When preprocessing the DWI image data and B0 image data by the data preprocessing unit (520), at least one of size adjustment, interpolation, normalization, and registration may be performed on the received DWI image data and B0 image data (however, the present invention is not necessarily limited thereto, and any preprocessing process necessary for applying an artificial intelligence model may be included and performed).

[0165] Combined image data is generated by combining the DWI image data and B0 image data preprocessed by the data combining unit (530) (S730). The combined image data is modeled by the DB0M classification modeling unit (540) using the pre-learned DB0M classification model provided by the DB0M classification model learning unit (560) to derive a DB0M inference value (S740).

[0166] At this time, when deriving a DB0M inference value using a DB0M classification model pre-learned by the DB0M classification modeling unit (540), the combined image data is input into an implemented deep neural network model using at least one of pre-learned CNN, 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception capable of processing image data, and the DB0M inference value can be derived through the inference process of the corresponding deep neural network model.

[0167] Thereafter, DB0M classification is performed based on the DB0M inference value derived by the DB0M classification unit (550) (S750). For example, when performing DB0M classification based on the derived DB0M inference value, the derived DB0M inference value is compared with a predetermined DB0M classification criterion, and if the resulting DB0M inference value is greater than the DB0M classification criterion, the DWI image and B0 image are classified as being identical, and if the DB0M inference value is equal to or less than the DB0M classification criterion, the DWI image and B0 image are classified as being inconsistent.

[0168] Another embodiment of the present invention described above can automatically classify whether or not a stroke patient has DB0M by inputting only DWI image data and B0 image data of a stroke patient through deep learning-based artificial intelligence modeling, so that DB0M classification, which is used as an important indicator for determining treatment indications for acute stroke patients, can be quickly performed without being influenced by the subjective judgment of each medical expert regarding the image to be read.

[0169]

[0170] FIG. 8 is a configuration diagram of a stroke diagnosis system based on DAM automatic classification according to another embodiment of the present invention.

[0171] Referring to FIG. 8, a DAM automatic classification system (800) for rapid stroke diagnosis according to another embodiment of the present invention (hereinafter referred to as the DAM automatic classification system) includes an ADC map generation unit (810), an image data input unit (820), a data preprocessing unit (830), a data combination unit (840), a DAM classification modeling unit (850), and a DAM classification unit (860). In addition, the DAM automatic classification system (800) according to one embodiment may include a DAM classification model learning unit (870).

[0172] According to another embodiment, the ADC map generation unit (810) of the DAM automatic classification system (800) generates an ADC map (also referred to as an ADC map image or ADC map image data) based on an ADC (Apparent Diffusion Coefficient) value derived by measuring signal attenuation due to diffusion from DWI (Diffusion-Weighted Imaging) image data to which different MRI (Magnetic Resonance Imaging) diffusion weights are applied. MRI refers to a diagnostic technology that uses a magnetic field to generate an image or photograph of a structure in the body and an image obtained thereby (in the following embodiments, the image is merely intended to emphasize the image itself, and the image data is merely intended to emphasize the data aspect for computer processing of the image, and in the following embodiments, the image and the image data are used interchangeably and are not used as different components of the invention).

[0173] ADC is an indicator that quantitatively represents the diffusion capacity of water molecules within a tissue. ADC reflects the microstructure and condition of the tissue and is affected by cell density, intracellular water mobility, etc. ADC is mm 2 It is expressed in units of / s and can be calculated based on [Mathematical Formula 1] based on the change in signal intensity obtained from the DWI image.

[0174]

[0175] Here, S0 is the signal intensity in the absence of MRI diffusion weighting (B0), S is the signal intensity at a specific MRI diffusion weighting, and B is a parameter that quantifies the diffusion sensitivity as a diffusion weighting in S.

[0176] An ADC map is a visual representation of the ADC values ​​in a specific tissue or area. The ADC map represents the ADC values ​​of each pixel (or voxel) across the entire scan area in color or intensity. The ADC map shows the spatial distribution of ADC values, making it easier to identify areas with pathological changes.

[0177] Areas of restricted diffusion are indicated by low ADC values ​​on ADC maps. In cases of acute cerebral infarction, areas of low ADC values ​​can be used to quantitatively assess damaged brain tissue. Because ADC maps quantitatively demonstrate the degree of restricted diffusion, they can assess not only the presence but also the degree of diffusion restriction (restricted intracellular water movement), making them crucial for stroke diagnosis.

[0178] ADC maps allow for quantitative assessment of diffusion restriction in areas identified on DWI images, such as B1000. Comparing B1000 and ADC map images allows for a more definitive assessment of whether diffusion restriction is occurring and whether the restriction is acute. This can be used to improve the accuracy of DAM assessment and to determine treatment indications for acute stroke patients.

[0179] According to another embodiment, the image data input unit (820) of the DAM automatic classification system (800) receives the generated ADC map and standard DWI image data to which MRI signal weights related to signal intensity exceeding a predetermined reference value are applied in an area where water molecule diffusion is restricted in acute stroke.

[0180] In the following examples, MRI diffusion weighted DWI image data is input, but it is not necessarily limited to the MRI method, and image data visualizing the internal structure and / or function of the brain can be input through direct or indirect methods such as computed tomography (CT) images, positron emission tomography (PET) images, and single photon emission computed tomography (SPECT) images.

[0181] In other embodiments below, the DWI image is image data obtained through a technique that measures the irregular movement of water molecules in a tissue, i.e., the diffusion velocity, and is generally an image processing method that can early measure cerebral infarction or lesions caused by cerebral vascular occlusion. By detecting such signal changes, various ischemic brain tissue damages such as the brain tissue damage area and damage volume can be imaged. At this time, the diffusion velocity is determined by the B value indicating the strength of the gradient magnetic field (the B value is a parameter that controls signal attenuation due to the diffusion of water molecules in the DWI image, and the higher the value, the greater the signal attenuation due to diffusion). In general, DWI images are known to be useful for quickly detecting acute or subacute cerebral infarction.

[0182] DWI (Diffusion-Weighted Imaging) is a technique for acquiring diffusion-weighted images, and various B values ​​can be used. In the following examples, DWI image data to which MRI signal weighting is applied, which is related to a signal intensity above a predetermined reference value in an area where water molecule diffusion is restricted in acute stroke, is referred to as standard DWI image data. In general, a B value of 1000 s / mm² (referred to as B1000) is known to be appropriate for emphasizing an area where water molecule diffusion is restricted, and for this reason, DWI image data to which MRI signal weighting of B1000, a B value widely used in stroke diagnosis, is applied can be used as standard DWI image data (however, various B values ​​can be applied for diffusion weighting when acquiring DWI images so that the movement of water molecules in tissues can be visualized from various angles by weighting diffusion differently, and DWI image data to which MRI signal weighting of various B values, such as B500, B1000, and B2000, is applied can be used as standard DWI image data depending on the characteristics of the disease, the purpose of diagnosis, the specifications of the MRI equipment, etc.).

[0183] According to another embodiment, the data preprocessing unit (830) of the DAM automatic classification system (800) can perform a preprocessing process so that the ADC map image data and standard DWI image data received from the image data input unit (820) can be used in the DAM classification modeling unit (850), which is an artificial intelligence modeling unit to be described later.

[0184] The preprocessing step performed by the data preprocessing unit (830) according to another embodiment may perform at least one of image size adjustment, interpolation, normalization, and image registration.

[0185] That is, the data preprocessing unit (830) according to another embodiment can adjust the size of each image (resize) to a predetermined resolution (for example, a resolution of 256 Х 256) when the resolution of the input ADC map image data and the standard DWI image data are not uniform, and then perform interpolation to configure each image data (B0 image data and DWI image data) into a plurality of images (for example, 36) (at this time, the nearest neighborhood method is used in the resize and interpolation process to enlarge or reduce the resolution and depth of the data format). After this, each image can be restricted to have a value between 0 and 1 through image-by-image normalization for all data formats.

[0186] In addition, the data preprocessing unit (830) according to another embodiment may perform denoising, intensity normalization, skull stripping, bias field correction, feature extraction, data augmentation, etc. (however, it is not necessarily limited thereto, and any preprocessing process necessary for applying an artificial intelligence model may be included and performed).

[0187] Noise removal, which can be performed as a preprocessing step, is a process of reducing noise in image data and improving the quality of image data by performing noise removal techniques such as Gaussian filter and non-local mean filter to remove noise that may be included in the process of acquiring image data.

[0188] Intensity normalization, which can be performed as a preprocessing step, is the process of standardizing image data intensity to reduce intensity differences between image data acquired at different times or from different image acquisition devices. Skull stripping, which can be performed as a preprocessing step, removes non-brain tissue, such as the skull and skin, from brain images, making it necessary when analyzing only brain tissue. Bias field correction, which can be performed as a preprocessing step, improves image uniformity by adjusting non-uniform signal intensities during image data acquisition to minimize the influence of bias fields.

[0189] Feature extraction, which can be performed as a preprocessing step, is the process of extracting useful features (e.g., information about edges, texture information, etc.) from an image before inputting it into a deep learning or machine learning model. These feature extractions can be utilized for analysis such as lesion classification and segmentation. Data augmentation, which can be performed as a preprocessing step, is the process of generating more training data by transforming existing image data through methods such as rotation, inversion, and resizing to improve the generalization ability of an artificial intelligence model.

[0190] According to another embodiment, the data combining unit (840) of the DAM automatic classification system (800) can combine the ADC map image data and the standard DWI image data preprocessed in the data preprocessing unit (830). That is, the data combining unit (840) according to one embodiment can combine (merge) the ADC map image data (for example, the dimension of (1, 36, 256, 256)) that has been resized and interpolated to match the same dimension and the standard DWI image data (for example, the dimension of (1, 36, 256, 256)) to generate combined image data (for example, combined image data of the dimension of (2, 36, 256, 256)) (the adjusted size and the interpolation method, etc., can be implemented in various ways depending on the implementation level).

[0191] Meanwhile, the order of data preprocessing and combining is not necessarily limited to performing preprocessing and then combining data. Depending on the implementation method, preprocessing may be performed on combined image data generated after combining ADC map image data and standard DWI image data.

[0192] According to another embodiment, the DAM classification modeling unit (850) of the DAM automatic classification system (800) performs inference to derive a DAM inference value by modeling (inferring) the combined image data input from the data combining unit (840) using a pre-learned DAM classification model.

[0193] As described above, the consistency between DWI images and FLAIR images is used as an important indicator for determining the timing of stroke treatment. According to one embodiment, the DAM classification modeling unit (850) can derive a DAM inference value from the input ADC map image and the standard DWI image using a DAM (Diffusion-Adc Mismatch) classification model that classifies the consistency between the pre-learned (pre-learned) ADC image and the standard DWI image.

[0194] The DAM classification model is a deep neural network model that is pre-trained using multiple training DWI images and training ADC images, with DFM set as a reference label.

[0195] In another embodiment, a DAM classification model used in the DAM classification modeling unit (850) may be a convolutional neural network (CNN), and a deep learning-based 3D convolutional neural network (CNN) such as 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception may be used. Such a 3D CNN model can process 3D data to recognize and learn complex patterns, to learn and recognize complex structures in medical image analysis, to be useful for capturing important structural features in high-resolution data such as medical images, and to extract features of images and solve classification or regression problems using multiple convolutional layers, and to improve the training process by optimizing the way information flows throughout the depth of the network (for example, DenseNet directly passes information between all layers, and ResNet uses a residual connection method that adds the input to the output).

[0196] A DAM classification modeling unit (850) according to another embodiment may use a 3D-DenseNet model as illustrated in FIG. 9. That is, as illustrated in FIG. 9, a DAM classification modeling unit (850) according to another embodiment may use a 3D DenseNet as a binary classification model for DAM, and at this time, a CNN network architecture may be implemented with a 4-time repeated dense block and a transition layer that uses combined image data, in which DWI image data and B0 image data are combined, as input.

[0197] According to another embodiment, the DAM classification modeling unit (850) may have a basic structure such as a dense block composed of two convolutional layers (for example, 1Х1Х1 and 3Х3Х3 convolutional layers), and this structure may be implemented so that the depth of the block increases by repeating. At this time, multiple (for example, 4) dense blocks of the CNN network may have various depths depending on the order, and may have two convolutional layers that are repeated 6, 12, 24, and 16 times respectively depending on the order of the blocks (the number of repetitions may be changed depending on the implementation), and the last linear layer may have multiple (for example, 2) nodes, and may be implemented so as to classify matching and mismatching cases depending on the scale of the value of each node.

[0198] At this time, in the DAM classification modeling unit (850) according to another embodiment, batch normalization (BN) and 3D-Conv (3D convolution) can be applied, and a stride of 2 (but not necessarily limited thereto) can be uniformly used in the convolution layer and pooling layer excluding the dense block and transition layers.

[0199] In general, the final output layer of the neural network can calculate the probability for each class (matched and unmatched) through a linear layer and finally using a softmax activation function, and finally perform classification (also called discrimination) based on the calculated probability value. In this way, the DAM classification modeling unit (850) according to one embodiment inputs the input combined image data into a 3D CNN model, and as shown in FIG. 9, it inputs the DAM inference value corresponding to the probability for each class (matched and unmatched) through a linear layer and finally using a softmax activation function. ) is derived.

[0200] According to another embodiment, the DAM classification unit (860) of the DAM automatic classification system (800) performs DAM classification based on the DAM inference value transmitted from the DAM classification modeling unit (850). That is, the DAM classification unit (860) according to another embodiment performs DAM classification based on the transmitted DAM inference value (as shown in FIG. 9). ) is classified as a matched case if it is greater than a given classification criterion (e.g. 0.7). ), if it is below the classification criterion, it is classified as a mismatch (unmatched case). ) can be done (depending on the implementation method, the classification method can be set differently for cases that match and do not match the classification criteria).

[0201] In this way, the DAM classification unit (860) according to another embodiment can classify DAM using only ADC map image data and standard DWI image data without FLAIR image data based on a pattern obtained in a learning process described later (i.e., the function of classifying whether a DWI image and an ADC image match can be performed using only the ADC map image and the standard DWI image).

[0202] A DAM automatic classification system (800) according to another embodiment may further include a DAM classification model learning unit (870). The DAM classification model learning unit (870) according to another embodiment may train a DAM classification model using learning data (at least one of a plurality of learning DWI images, learning FLAIR images, and learning ADC map images).

[0203] In addition, the DAM classification model learning unit (170) according to another embodiment can pre-train (train) the DAM classification model by creating a pre-learned (pre-trained) model using a method that precedes pretext training, such as self-supervised learning using a collection of natural images or MRI images such as brain DWI images and ADC images without lesions, and then fine-tuning the model.

[0204] FLAIR is one of the MRI techniques. It attenuates the signal of body fluids such as cerebrospinal fluid (CSF) by a set value of inversion time (TI), which is a variation of the inversion recovery sequence, to make CSF appear darker in the captured image, thereby making the difference with contrasting lesions or tissues more prominent (inversion time refers to the time required to minimize the signal of a specific tissue in an MRI scan, and by adjusting this time when acquiring a FLAIR image, the signal of CSF can be eliminated, resulting in more distinctly emphasized pathological changes in the brain).

[0205] This attenuation of CSF signals allows for more clearly highlighting brain lesions or tissue changes, and enhances the contrast between white and gray matter, as well as between lesions and normal tissue, allowing for a more accurate assessment of various brain pathological conditions. FLAIR imaging is generally known to be useful for assessing the temporal course of cerebral infarction.

[0206] Another method for learning and generating a DAM classification model in a DAM classification model learning unit (870) according to another embodiment is as follows. In a medical field, by examining learning ADC images and / or learning standard DWI image data and learning ADC map images for a number of clinical patients, the presence or absence of DAM is determined, and the resulting generated ground truth data can be used as an accurate label of DAM and as a training dataset for the DAM classification model of the present invention.

[0207] Through this process, the DAM classification model learning unit (870) according to one embodiment can generate a DAM classification model capable of recognizing a pattern between DAM labels associated with at least one of DWI image data, FLAIR image data, and ADC map image data. That is, the DAM classification model learning unit (870) according to one embodiment can learn the characteristics of ADC map image data and standard DWI image data by utilizing a machine learning or deep learning algorithm, and build a model capable of classifying whether or not there is DAM even without FLAIR image data.

[0208] According to another embodiment, the DAM classification model learning unit (870) can perform processes related to de-identification of image data and data protection to safely manage medical image data and protect patient personal information.

[0209] In another embodiment, the image data de-identification process performed in the DAM classification model learning unit (870) may include a screening for collected data process for selecting and reviewing available medical image data, a removing personal information from image files process for downloading image files of the selected data and removing the patient's personally identifiable information from the image files to protect the privacy of the individual, and a labeling and subject conversion process for assigning a new label to the de-identified data or converting it into a new subject for patient identification.

[0210] In another embodiment, the de-identified patient basic information processing process performed in the DAM classification model learning unit (870) may perform a de-identified patient basic information processing process that removes personally identifiable elements from the patient's basic information to make it impossible to identify an individual, which is de-identified personal information, and anonymized information, which is information processed so that an individual can no longer be identified, and a risk of re-identification process that considers the risk of patient information being re-identified even during the de-identification process.

[0211] In another embodiment, the DAM classification model learning unit (870) may perform an operator management process for managing operators who process data, a user registration process for registering users who will use the data processing system, an inspector / operator assignment process for allocating personnel to supervise and operate each data processing process, a data processing process for processing de-identified data, annotation process for annotating data or adding metadata, and a first inspection and second inspection sampling process for checking whether the data processing and de-identification processes have been performed appropriately.

[0212] In this way, the DAM classification model learning unit (870) according to another embodiment can perform various processes for protecting personal information and ensuring data security as described above, and, in contrast, a separate configuration module (not shown) can perform various processes for protecting personal information and ensuring data security for data used in the learning and inference processes.

[0213] Meanwhile, in another embodiment, the DAM classification model learning unit (870) can automatically or manually label data based on the characteristics of different tissues or lesions exhibiting different signal intensities in medical images, such as MRI or CT. This allows for the classification of specific tissues or lesions and the extraction of information necessary for medical diagnosis (in this case, signal intensity discrepancies refer to differences compared to a specific standard or expected value, and data points can be assigned appropriate labels based on these differences).

[0214] According to another embodiment, the DAM classification model learning unit (870) may utilize a cross-validation method such as k-fold cross-validation. In this case, the data set for learning is divided into multiple partitions (folds), and each partition may be used for training, tuning, and testing different portions of the data set (Training Set), thereby enhancing the generalization ability of the DAM classification model by allowing the model to learn and be tested on various portions of the data.

[0215] According to another embodiment, the DAM classification model learning unit (870) divides a dataset including a plurality of learners (stroke patients) into N segments (folds), and each segment is configured to operate as a training set (N-1) and a tuning and test set (1), and can learn using the dataset. The training set is a dataset used to learn the artificial intelligence model, the tuning set is a dataset used to adjust the hyperparameters of the artificial intelligence model, and the test set is a dataset used to evaluate how well the artificial intelligence model generalizes.

[0216] Specifically, when splitting the dataset into training, training, and test sets, a similar proportion of labels can be maintained in each set to maintain distributional consistency, and the predicted probabilities on the test set of the DAM classifier derived from each split are aggregated to form a cohort consisting of a large number of unique trainees, and performance metrics are applied to this aggregated data to provide an overall evaluation of the model's predictive ability. This comprehensive approach can emphasize the robustness of the performance evaluation by reflecting the model's generalizability to new, unseen data.

[0217] Meanwhile, the DAM classification model learning unit (870) according to another embodiment may use other dataset configuration methods that can be used for learning an artificial intelligence model in addition to the aforementioned method, such as a simple holdout method, a stratified k-fold cross-validation, a leave-p-out cross-validation, a time-based split, bootstrap, and Monte Carlo cross-validation (however, the present invention is not necessarily limited thereto, and dataset configuration methods for learning an artificial intelligence model that are not mentioned may be used). The DAM classification model learning unit (870) according to one embodiment may appropriately select a dataset configuration method according to the characteristics of the data, the complexity of the model, the precision requirements of the evaluation, and the like so that the model can work well on real-world data.

[0218] Hereinafter, a DAM automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention based on the above description will be described.

[0219]

[0220] FIG. 10 is an example of a DAM automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention, which is an example of a DAM automatic classification method for rapid stroke diagnosis using a DAM automatic classification device for rapid stroke diagnosis according to another embodiment of the present invention of FIG. 8.

[0221] Referring to FIG. 10, a DAM automatic classification method for rapid stroke diagnosis according to another embodiment of the present invention includes the steps of receiving standard DWI image data and ADC map to which MRI signal weights related to signal intensity exceeding a predetermined reference value in an area where water molecule diffusion is restricted in acute stroke are applied, performing preprocessing on the input standard DWI image data and ADC map, generating combined image data by combining the preprocessed standard DWI image data and ADC map, deriving a DAM inference value using the generated combined image data and a pre-learned DAM (Diffusion-ADC Mismatch) classification model, and performing DAM classification based on the derived DAM inference value.

[0222] Hereinafter, each step of the DAM automatic classification method for rapid stroke diagnosis of Fig. 10 will be described in detail. First, standard DWI image data and ADC maps with MRI signal weights applied that are related to a signal intensity higher than a predetermined standard value in an area where water molecule diffusion is restricted in acute stroke are input (received) by the image data input unit (120) (S1010).

[0223] At this time, the ADC map can be generated based on the ADC value derived by measuring signal attenuation due to diffusion from DWI image data to which different MRI diffusion weights are applied.

[0224] Preprocessing is performed on standard DWI image data and ADC map image data input by the data preprocessing unit (830) (S1020).

[0225] When performing preprocessing on standard DWI image data and ADC map image data by the data preprocessing unit (830), at least one of size adjustment, interpolation, normalization, and registration can be performed on the received standard DWI image data and ADC map image data.

[0226] Standard DWI image data and ADC map image data preprocessed by the data combining unit (840) are combined to generate combined image data (S1030).

[0227] The combined image data generated by the DAM classification modeling unit (850) and the pre-learned DAM (Diffusion-ADC Mismatch) classification model are used to model (infer) and derive a DAM inference value (S1040).

[0228] At this time, when deriving a DAM inference value using a DAM classification model pre-learned by the DAM classification modeling unit (850), the combined image data is input into an implemented deep neural network model that uses at least one of CNN, 3D-DenseNet, 3D-ResNet, 3D-U-Net, 3D-VGG, and 3D-Inception that can process pre-learned (pre-learned) image data, and the DAM inference value can be derived through the inference process of the corresponding deep neural network model.

[0229] Afterwards, DAM classification is performed based on the DAM inference value derived by the DAM classification unit (860) (S1050). When performing DAM classification based on the DAM inference value derived by the DAM classification unit (860), the derived DAM inference value is compared with a predetermined DAM classification criterion, and as a result, if the DAM inference value is greater than the DAM classification criterion, it is classified as a DAM match, and if the DAM inference value is equal to or less than the DAM classification criterion, it can be classified as a DAM mismatch (i.e., the function of classifying whether or not a DWI image matches an ADC image can be performed with only an ADC map image and a standard DWI image).

[0230] Another embodiment of the present invention described above can automatically and quickly classify whether or not a stroke patient has DAM by inputting only standard DWI image data and ADC map image data of a stroke patient through deep learning-based artificial intelligence modeling, thereby enabling a quick determination of treatment indications for acute stroke patients without being influenced by the subjective judgment of each medical expert regarding the image to be read.

[0231]

[0232] The above-described embodiments can be implemented using various types of computing means including one or more processors, memories, and storage means. In addition, a network interface connected to a wired or wireless network can be included. The processor can be a central processing unit or a semiconductor device that executes processing instructions stored in a memory and / or storage unit. The memory and the storage unit can include volatile storage media or non-volatile storage media. For example, the memory can include ROM and RAM. Accordingly, embodiments of the present invention can be implemented as a method implemented by a computer or as a non-transitory computer-readable medium having computer-executable instructions stored on the computer. In one embodiment of the present invention, when executed by a processor, the computer-readable instructions can perform a method according to at least one aspect of the present invention.

[0233] Although the present invention has been described with reference to the illustrated embodiments as above, these are merely exemplary, and it will be apparent to those skilled in the art to which the present invention pertains that various modifications, changes, and equivalent various other embodiments are possible without departing from the spirit and scope of the present invention. For example, the data preprocessing unit (130) and the data combining unit (140) may be implemented by integrating them into a single module, or may be implemented by dividing them into two or more devices. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

[0234] The present invention can be used in a stroke diagnosis system, etc.

Claims

1. Image data input unit that receives DWI (Diffusion-Weighted Imaging) images and FLAIR (Fluid-Attenuated Inversion Recovery) images; A data combining unit that combines the input DWI image and FLAIR image to generate combined image data; A DFM classification modeling unit that derives a DFM inference value using the generated combined image data using a DFM (Diffusion-FLAIR Mismatch) classification model; and A DFM automatic classification system for stroke diagnosis, comprising a DFM classification unit that performs DFM classification based on the derived DFM inference values.

2. In paragraph 1, A DFM automatic classification system for stroke diagnosis further comprising a preprocessing unit for preprocessing the input DWI image and FLAIR image.

3. In paragraph 2, A DFM automatic classification system for stroke diagnosis, wherein the preprocessing unit performs at least one of resizing, interpolation, normalization, and registration on the input DWI image and FLAIR image.

4. In paragraph 1, The above DFM classification model is a deep neural network model pre-trained using a training DWI image collection and a training FLAIR image collection, and is a DFM automatic classification system for stroke diagnosis.

5. In paragraph 4, The above pre-learned deep neural network model is a DFM automatic classification system for stroke diagnosis implemented using a convolutional neural network (CNN).

6. In paragraph 1, The above DFM classification section By comparing the above derived DFM inference value with the given DFM classification criteria, A DFM automatic classification system for stroke diagnosis, wherein if the DFM inference value is greater than the DFM classification criterion, the DWI image and the FLAIR image are classified as matching, and if the DFM inference value is equal to or less than the DFM classification criterion, the DWI image and the FLAIR image are classified as mismatching.

7. A step of performing preprocessing on the received DWI (Diffusion-Weighted Imaging) image and FLAIR (Fluid-Attenuated Inversion Recovery) image; A step of combining the preprocessed DWI image and FLAIR image to generate combined image data; A step of deriving a DFM inference value based on the generated combined image data and a pre-trained DFM (Diffusion-FLAIR Mismatch) classification model; and A DFM automatic classification method for stroke diagnosis, comprising a step of performing DFM classification based on the derived DFM inference value.

8. In paragraph 7, The step of performing preprocessing on the received DWI image and FLAIR image is A DFM automatic classification method for stroke diagnosis, which performs at least one of resizing, interpolation, normalization, and registration on the received DWI image and FLAIR image.

9. In paragraph 7, The step of deriving a DFM inference value based on the generated combined image data and the pre-trained DFM (Diffusion-FLAIR Mismatch) classification model is as follows. A step of inputting the combined image data into a pre-trained deep neural network model using a training DWI image collection and a training FLAIR image collection; and A DFM automatic classification method for stroke diagnosis, comprising a step of deriving a DFM inference value through the inference process of the above deep neural network model.

10. In paragraph 9, The above deep neural network model is a DFM automatic classification method for stroke diagnosis implemented using a CNN (Convolutional Neural Network).

11. In paragraph 7, The step of performing DFM classification based on the above derived DFM inference value is A step of comparing the above derived DFM inference value with a predetermined DFM classification criterion; A step of classifying the DWI image and the FLAIR image as matching if the DFM inference value is greater than the DFM classification criterion; and A DFM automatic classification method for stroke diagnosis, comprising a step of classifying the DWI image and the FLAIR image as inconsistencies if the DFM inference value is equal to or less than the DFM classification criterion. 12.B0 Image data input unit for receiving image data and MRI diffusion-weighted DWI (Diffusion-Weighted Imaging) image data; A data combining unit that combines the input B0 image data and MRI diffusion weighted DWI image data to generate combined image data; A DB0M classification modeling unit that receives the generated combined image data and derives a DB0M inference value using a pre-trained DB0M (Diffusion-B0 Mismatch) classification model; and A DB0M automatic classification system for stroke diagnosis, comprising a DB0M classification unit that performs DB0M classification based on the derived DB0M inference value.

13. In paragraph 12, A DB0M automatic classification system for stroke diagnosis further comprising a preprocessing unit for preprocessing the input B0 image data and MRI diffusion weighted DWI image data.

14. In paragraph 13, A DB0M automatic classification system for stroke diagnosis, wherein the above preprocessing unit performs at least one of size adjustment, interpolation, normalization, and registration on the input B0 image data and MRI diffusion-weighted DWI image data.

15. In paragraph 12, The above pre-trained DB0M classification model is a DB0M automatic classification system for stroke diagnosis, which is a deep neural network model pre-trained using multiple training DWI images and training B0 images, with DFM set as a reference label.

16. In paragraph 15, The above pre-learned deep neural network model is a DB0M automatic classification system for stroke diagnosis implemented using a convolutional neural network (CNN).

17. In paragraph 12, The above DB0M classification unit compares the derived DB0M inference value with a predetermined DB0M classification criterion. A DB0M automatic classification system for stroke diagnosis, which classifies a DB0M match if the DB0M inference value is greater than the DB0M classification criterion, and classifies a DB0M mismatch if the DB0M inference value is equal to or less than the DB0M classification criterion. 18.B0 Step of receiving image data and MRI diffusion-weighted DWI (Diffusion-Weighted Imaging) image data; A step of performing preprocessing on the input B0 image data and DWI image data; A step of combining the preprocessed B0 image data and DWI image data to generate combined image data; A step of deriving a DB0M inference value using the generated combined image data and a pre-learned DB0M (Diffusion-B0 Mismatch) classification model; and A DB0M automatic classification method for stroke diagnosis, comprising a step of performing DB0M classification based on the derived DB0M inference value.

19. In paragraph 18, The step of performing preprocessing on the above input B0 image data and DWI image data is A DB0M automatic classification method for stroke diagnosis, which performs at least one of resizing, interpolation, normalization, and registration on the input B0 image data and DWI image data.

20. In paragraph 18, The step of deriving a DB0M inference value using the above-generated combined image data and the pre-learned DB0M (Diffusion-B0 Mismatch) classification model is A step of inputting the combined image data into a pre-trained deep neural network model using at least one of a plurality of training DWI images and training B0 images; and A DB0M automatic classification method for stroke diagnosis, comprising a step of deriving a DB0M inference value through the inference process of the above deep neural network model.

21. In paragraph 20, The above deep neural network model is a DB0M automatic classification method for stroke diagnosis implemented using a CNN (Convolutional Neural Network).

22. In paragraph 18, The step of performing DB0M classification based on the above derived DB0M inference value is A step of comparing the above derived DB0M inference value with a predetermined DB0M classification criterion; A step of classifying as DB0M match if the above DB0M inference value is greater than the above DB0M classification criterion; and A DB0M automatic classification method for stroke diagnosis, comprising a step of classifying a DB0M mismatch if the DB0M inference value is equal to or less than the DB0M classification criterion.

23. An ADC map generation unit that generates ADC map image data based on an ADC (Apparent Diffusion Coefficient) value derived by measuring signal attenuation due to diffusion from DWI (Diffusion-Weighted Imaging) image data to which different MRI (Magnetic Resonance Imaging) diffusion weights are applied; An image data input unit that receives the generated ADC map image data and standard DWI image data to which MRI signal weighting is applied in relation to a signal intensity exceeding a predetermined reference value in an area where water molecule diffusion is restricted in acute stroke; A data combining unit that combines the input standard DWI image data and ADC map image data to generate combined image data; A DAM classification modeling unit that receives the generated combined image data and derives a DAM inference value using a DAM (Diffusion-ADC Mismatch) classification model; and A DAM automatic classification system for stroke diagnosis, comprising a DAM classification unit that performs DAM classification based on the derived DAM inference value.

24. In paragraph 23, A DAM automatic classification system for stroke diagnosis, further comprising a preprocessing unit for preprocessing the input standard DWI image data and ADC map image data.

25. In paragraph 24, A DAM automatic classification system for stroke diagnosis, wherein the above preprocessing unit performs at least one of size adjustment, interpolation, normalization, and registration on the input standard DWI image data and ADC map image data.

26. In paragraph 23, The above DAM classification model is a DAM automatic classification system for stroke diagnosis, which is a deep neural network model pre-trained using multiple training DWI images and training ADC images, and sets DFM as a reference label.

27. In paragraph 26, The above pre-learned deep neural network model is a DAM automatic classification system for stroke diagnosis implemented using a convolutional neural network (CNN).

28. In paragraph 23, The above DAM classification section By comparing the above derived DAM inference value with the given DAM classification criteria, A DAM automatic classification system for stroke diagnosis, which classifies a case as a DAM match if the DAM inference value is greater than the DAM classification criterion, and classifies a case as a DAM mismatch if the DAM inference value is equal to or less than the DAM classification criterion.

29. A step of inputting standard DWI image data and ADC map image data to which MRI signal weighting is applied in relation to a signal intensity exceeding a predetermined standard value in an area where water molecule diffusion is restricted in acute stroke; A step of performing preprocessing on the input standard DWI image data and ADC map image data; A step of combining the preprocessed standard DWI image data and ADC map image data to generate combined image data; A step of deriving a DAM inference value using the generated combined image data and a pre-trained DAM (Diffusion-ADC Mismatch) classification model; and A method for automatically classifying DAM for stroke diagnosis, comprising a step of performing DAM classification based on the derived DAM inference value.

30. In paragraph 29, The above ADC map image data is a DAM automatic classification method for stroke diagnosis, which is generated based on an ADC value derived by measuring signal attenuation due to diffusion from DWI image data to which different MRI diffusion weights are applied.

31. In paragraph 29, The step of performing preprocessing on the above input standard DWI image data and ADC map image data is; A DAM automatic classification method for stroke diagnosis, which performs at least one of resizing, interpolation, normalization, and registration on the input standard DWI image data and ADC map image data.

32. In paragraph 29, The step of deriving a DAM inference value using the above-mentioned generated combined data and the pre-learned DAM classification model is A step of inputting the combined image data into a pre-trained deep neural network model using a plurality of training DWI images and training ADC images; and A DAM automatic classification method for stroke diagnosis, comprising a step of deriving the DAM inference value through the inference process of the deep neural network model.

33. In paragraph 32, The above deep neural network model is a DAM automatic classification method for stroke diagnosis implemented using a CNN (Convolutional Neural Network).

34. In paragraph 23, The step of performing DAM classification based on the above derived DAM inference value is A step of comparing the above derived DAM inference value with a predetermined DAM classification criterion; A step of classifying as DAM matching if the above DAM inference value is greater than the above DAM classification criterion; and A DAM automatic classification method for stroke diagnosis, comprising a step of classifying a DAM mismatch if the DAM inference value is equal to or less than the DAM classification criterion.

Images